Unraid management agent

Unraid Management Agent

The Unraid Management Agent is a plugin that runs on an Unraid server and exposes its full state and control surface over three interfaces:

Interface	Endpoint	Use for
MCP (Model Context Protocol)	http://<unraid-ip>:8043/mcp	AI agents — 125 tools, 5 resources, 6 prompts
REST API	http://<unraid-ip>:8043/api/v1	Scripts, Custom GPT Actions, integrations
WebSocket	ws://<unraid-ip>:8043/api/v1/ws	Live event streaming (near real-time)

Core principle: prefer the MCP tools when acting as an AI agent — they are purpose-built, validated, and named for discovery. Fall back to the REST API only for clients that cannot speak MCP (e.g. ChatGPT Custom GPTs). Never invent tool names or paths; the real catalogs are in the reference files below.

Decision Workflow

1. Connect

If the client is not yet connected, read references/connection.md. In short:

• Streamable HTTP (remote/LAN): point the client at http://<unraid-ip>:8043/mcp.
• STDIO (local on the server): run unraid-management-agent mcp-stdio.
• The agent has no authentication by default — it is designed for a trusted LAN or behind a VPN/reverse proxy. Do not assume an API key exists unless the user says so.

2. Read before you write

Always establish current state with a read-only tool before taking action:

• System snapshot → get_system_info, or get_diagnostic_summary for a broad view
• Array → get_array_status; Disks → list_disks / get_disk_info
• Containers → list_containers / search_containers / get_container_info
• VMs → list_vms / search_vms / get_vm_info

Read-only tools are marked ReadOnlyHint: true. The full catalog is in references/mcp-tools.md.

3. Pick the narrowest tool

• Looking for one container/VM among many? Use search_containers / search_vms, not a full list_ dump.
• Want a specific subsystem? Use the dedicated getter (get_ups_status, get_gpu_metrics, get_zfs_pools, …) rather than scraping get_diagnostic_summary.
• Need history/metrics over time? Use query_metric_history, get_parity_history, get_alert_history, get_health_check_history.

4. Confirm destructive actions

Control tools (ReadOnlyHint: false) change the system. The high-risk ones require confirm=true and must be confirmed with the user first:

Tool	Effect
array_action (stop)	Makes all array data inaccessible
system_reboot / system_shutdown	Restarts / powers off the whole server
execute_user_script / run_runbook	Runs arbitrary user-defined actions
delete_vm_snapshot / restore_vm_snapshot	Irreversible VM state change
delete_alert_rule / delete_health_check	Removes configuration

Always state what will happen and wait for explicit user approval before sending confirm=true. See references/mcp-tools.md for the full read/write breakdown.

5. Use live data correctly

For continuously changing values (CPU, container state), do not poll the REST API in a tight loop. Use the MCP resources (references/diagnostics.md) or the WebSocket stream. One MCP tool call returns the latest cached value already.

6. Diagnose with prompts, not ad-hoc chains

The agent ships 6 built-in diagnostic prompts that orchestrate the right tools for common investigations (disk health, performance, maintenance, array state, overview, troubleshooting). Prefer them over reinventing the analysis. See references/diagnostics.md.

---

Critical Anti-Patterns

Anti-pattern	Use instead	Why
Inventing tool names (docker_restart, getArray)	The exact names in references/mcp-tools.md	Wrong names just fail
Stopping the array / rebooting without confirmation	Read state, explain impact, then confirm=true	Data loss / downtime
Polling GET /api/v1/system in a loop	MCP resources or WebSocket ws://…/api/v1/ws	The agent already caches; polling wastes CPU
list_containers then filtering in-context	search_containers / get_container_info	Less context, faster, more accurate
Hand-rolling a disk-health analysis	diagnose_disk_health prompt	Built-in prompt covers SMART, temps, wear, errors
Assuming an API key / auth header	Treat as unauthenticated LAN/VPN unless told otherwise	Default deploy has no auth
Using REST from an MCP-capable client	The MCP tools	Validated args, richer results, fewer round-trips
Editing Unraid config files directly	The agent's control tools / settings endpoints	Tools match the WebUI's own logic and validation

---

Reference Files

Read these when you need detail. Keep SKILL.md itself in context; pull references on demand.

File	When to read
references/connection.md	Connecting any client (Claude Code/Desktop, claude.ai, Cursor, Copilot, Gemini CLI, ChatGPT) via MCP HTTP/STDIO; troubleshooting connectivity
references/mcp-tools.md	The full catalog of all 125 MCP tools grouped by category, with read/write (destructive) flags and arguments
references/diagnostics.md	The 6 diagnostic prompts and 5 real-time resources — what each analyses and when to use it
references/rest-api.md	REST API surface for non-MCP clients (Custom GPT Actions, scripts): base path, key endpoints, path/body conventions
references/workflows.md	Natural-language request → tool/endpoint mappings and end-to-end examples (monitoring, container/VM/array control, maintenance)

Capabilities at a Glance

• Monitoring (read-only): system, array, disks + SMART, shares, Docker (containers, networks, stats, logs, update checks), VMs + snapshots, GPU, UPS, NUT, ZFS (pools/datasets/snapshots/ARC), network + access URLs, hardware, notifications, parity history, logs, collectors, settings, unassigned devices, remote SMB/NFS shares.
• Control: container & VM lifecycle, array start/stop, parity check start/stop/pause/resume, disk spin up/down, container & plugin updates, VM snapshot create/delete/restore, VM clone, service start/stop/restart, user scripts, collector enable/disable + intervals, reboot/shutdown.
• Observability & automation: alert rules (expr-lang) + templates, health checks (HTTP/TCP/container), remediation runbooks, an optional autonomous agent, fan control (profiles/curves), CPU governor, and system tuning (turbo boost, disk cache, inotify).